Due to the increasing popularity of clinical decision support systems, the problem of personalized drug dose identi cation becomes more relevant and substantial. In this paper, the authors introduce a data-driven model designed to operate in this case. Current work comprises general problem formulation, description of data and its preprocessing steps, model design overview, its rst stage model tuning and training, evaluation metrics used to estimate the quality, achieved values.
Currently, decision support systems that consider patient characteristics are
gaining more popularity and impact on the process of treatment [
ITR was studied for various diseases, such as oncology [
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). ? The reported study was funded by RFBR according to the research project #18-37-00441.
This task of obtaining ITR was being solved in speci c areas with such
approaches as the Q-learning [
In this paper, the authors consider a supervised learning approach to obtain ITR for personalized combined drug therapy. 2
Given the vector Xj (ti) of patient pro le j in the moment of time ti before treatment as:
Xj (ti) = nx(jh)(ti)o Obtain the set of antihypertensive therapy:
Yj (Xj (ti)) = n(yj(k;1); yj(k;2))o; (1) (2) where j = 1; m; i = 1; n; h = 1; p; k = 1; q and each drug yjk is a vector containing a drug International Nonproprietary Name (INN) and optimal daily dosage. The model, which authors propose in current paper, is designed to consist of three data-driven submodels: the rst model receives a vector of patient features as an input and predicts an optimal drugs count (nopt), the second model extends the result specifying drug INNs nyj(k;1)onopt and the third model de nes the desired daily dosages of each drug INN nyj(k;k2=) o1kn=op1t resulting with n(yj(k;1); yj(k;2))okn=op1t . 3
The data used in this study were collected from 2010 to 2015, depersonalized and provided by Almazov National Medical Research Centre.
In the work, 16 features are grouped into a vector describing the patient pro le. They are the following: age, sex, body mass index (BMI), systolic and diastolic blood pressure before treatment, smoking status, impaired glucose tolerance (IGT), left ventricular hypertrophy (LVH), chronic heart failure (CHF), ischemic heart disease (IHD), dyslipidemia, diabetes, microalbuminuria, cardiovascular diseases (CVD) among relatives, chronic kidney disease stage (CKD), and combination of concomitant drugs.
In the authors' previous study, eight patient clusters were been identi cation based on the feature tuples. The patient groups accept the clinical interpretation. As time passes, a patient pro le will be changing due to disease development and ageing patient. It means that the same patient can belong to various groups depending on the disease dynamics at di erent points in time. To probabilistic model the arterial hypertension development in a patient, we use a M arkov chain of transition from one cluster to another cluster. Therefore, the dynamic model of the hypertensive patient transitions process from cluster i (i = 1; 8) to cluster j (j = 1; 8) is presented as a graph, where the nodes are clusters and the edges are transition probabilities Pij (Fig. 1). It's noted, the probabilities Pij are determined in a way that the transition probabilities sum is equal to one. Also, the patient condition will be able to remain the same then the patient will transition to the same cluster. These transitions are presented as a loop on between groups graph in Fig. 1. However, the certain clusters are incompatible for relative transitions, in particular, due to gender characteristics and/or the chronic concomitant diseases.
0.22 Cluster 1
Additionally, training dataset contains an outcome eld as a result of the
one-month treatment process. This value accepts various ways to be set. In this
particular research, it is based on clinical guidelines and points out if systolic and
diastolic blood pressure levels have reached the target values less than 140/90
mm Hg [
Since the task is to predict the optimal therapy for each patient based on his/her features vector, such elds as drug names and daily dosages have to be in the training data. So that, data was merged (on patient ID and outpatient visit date) with records contained medical prescriptions for these patients. As long as medical texts are written in natural language, they require additional processing to distinguish desired information and ll these elds.
Due to the lack of training dataset, the processing task was resolved in this research with a sequence of regular expressions and extracting rules implementing so-called rule-based natural language processing (NLP). Below is the overall pipeline that was applied to each outpatient visit: 1. Split the medical prescriptions eld into substrings with `nn' (newline) delimiter. In the provided data most of such substrings contain, if they do, only one prescription of the drug. 2. Find all entries of drugs in each substring using the dictionary prepared by authors. This dictionary includes drug brand-names, their di erent writing options that may show up in a natural language text, INNs, and pharmacological classes. 3. The substring may have no medical prescriptions because it has general guidance, referral to laboratory testing, etc. Such substring is not involved in further processing. 4. If substring contains several drug brand-names, then distinguish them as an alternative or as a combination. Patterns, which are used in this step, include checking: their INNs { the same indicates the alternative, their location in string boundaries, and the presence of `and' symbols, commas between them, conjunctions. 5. Check the presence of words that mean cancellation, dosages and frequency indicators. Substrings that don't have dosages and frequency can be involved in the dataset with lling missing values using appropriate machine learning algorithms in further preprocessing. 6. Extract dosages using regular expressions with measurement units, frequencies { using regular expressions with parts of the day patterns. 7. Aggregate INNs of all extracted drugs in the INN eld, calculate their daily dosages and write them in the Dosage eld using speci ed delimiter (in this research `j'). 5
This paper is aimed to describe the rst submodel in detail, which is proposed to be an extension of a treatment outcome classi er. In the cycle, it concatenates the vector of the patient features with every possible drug count g = 1; r and utilizes the classi er to predict the probabilities of treatment ine ectiveness (negative class, 0) and e ectiveness (positive class, 1) denoted as f(p(0); p(1))gg.
The combination with the maximum outcome probability is assumed to contain an optimal number of drugs.
nopt = arg max (p(0); p(1))g p(1) (3)
In the current Python implementation, several scikit-learn classi ers (library version 0.21.3) were trained, tuned and evaluated, including C-Support Vector Classi er (SVC), random forest classi er (RF), Multi-layer Perceptron (MLP), classi er as well as LightGBM (library version 2.3.0).
Hyper-parameters estimation using cross-validation led to the following values (parameters not mentioned below are expected to have the default values for the speci ed library version): { SVC: radial basis function (RBF) kernel, balanced class weights, enabled probability estimation, kernel coe cient = 0:04, penalty parameter of the error term C = 8:0. { Random forest classi er: entropy criterion, balanced class weights, with 150 trees in the forest of maximum depth 2 and 10 minimum samples to split. { MLP: stochastic gradient descent (sgd) solver, invscaling learning rate, maximum number of iterations is 20, one hidden layer with 76 neurons, regularization term is 0.2, using Nesterov's momentum, shu e samples in each iteration. { LightGBM: random forest boosting type, balanced class weights, bagging frequency is 1, bagging fraction is 0.9, learning rate is 0.01, number of estimators is 110. 6
After the preprocessing step the dataset containing 4521 records were divided into three datasets: training, validation, and test in a ratio of 0.63:0.27:0.1 respectively. It was decided to consider both Sensitivity (Recall) and Speci city while estimating the treatment outcome classi ers parameters. This decision is based on the requirement to e ciently identify both ine ective and e ective treatment and use it in revealing the optimal drug amount.
The rst two datasets were used in 5 splits with 7 repeats cross-validation, results of which are presented in Table 1. Table 2 gives the results of classi ers quality evaluation on the third (test) dataset. Although Sensitivity and Specicity were considered as target metrics, the tables additionally include values of such metrics as Accuracy, Precision, F1 score and ROC AUC (Receiver Operating Characteristic Area Under ROC Curve).
As can be seen from the tables that all classi ers avoided over tting. MLP classi er performed worse than the others, which showed comparatively close results. The random forest classi er is assumed to perform the most optimal way.
The most important patient features and their importances with more than 1% impact returned by random forest classi er are presented in Fig. 2. They are calculated as the impurity decrease from each feature: the reduce in node impurity weighted by the probability of reaching the node. It can be seen that the drug count provides around 4.0% of the total decision.
However, authors assume that using a bigger training dataset will improve the results and the impact of the drug count feature, which is now considered to be not su cient enough to reliably separate the e ective therapy from the ine ective.
As a result of this work, the model that predicts the optimal antihypertensive drug count was presented, implemented, trained, and evaluated. This model is a part of the proposed general model predicting the optimal antihypertensive drug dosages based on the patient features.
Future works of this research include: { preparation and preprocessing of new data collected from 2016 to 2019 that will be provided by Almazov National Medical Research Centre; { further training and parameters tuning of additional classi ers predicting the optimal amount of prescription drugs for a patient with arterial hypertension; { development of the data-driven model predicting the most e ective individual antihypertensive therapy including drug INNs and daily dosages.
The reported study was funded by RFBR according to the research project #18-37-00441.